Text-independent speaker identification using vocal tract length normalization for building universal background model

A. K. Sarkar, S. Umesh, S. P. RathInterspeech 2009

TL;DR

This paper proposes a novel approach to speaker identification by using Vocal Tract Length Normalization (VTLN) to build a more compact Universal Background Model (UBM). This leads to better-separated speaker-specific models and ultimately lower identification error rates compared to traditional GMM-UBM methods.

What Problem Does This Paper Address?

Speaker identification systems aim to determine who is speaking based on their voice. A common approach involves training a Universal Background Model (UBM) on a large, diverse dataset of speech. This UBM acts as a statistical representation of general speech characteristics. However, UBMs trained this way can be "coarse" due to inherent variations in speech, such as differences in vocal tract length (VTL) between speakers. When speaker-specific models are adapted from such a coarse UBM, they can overlap significantly in the acoustic space, making it harder to distinguish between speakers and leading to higher error rates. The authors aim to address this by creating a more "compact" UBM that reduces inter-speaker variability, thereby improving the separability of adapted speaker models.

Methods

The core innovation here is the application of Vocal Tract Length Normalization (VTLN) during the UBM training phase. VTLN is a technique that warps the frequency spectrum of speech to reduce variations caused by differences in vocal tract length. Instead of training the UBM on raw speech features, the authors first apply VTLN to the training utterances. This results in a VTLN-UBM, which is hypothesized to be more compact than a conventional GMM-UBM.

The process for building the VTLN-UBM involves:

  1. Estimating Warp Factors: For each training utterance, warp factors are estimated by maximizing the likelihood of the warped utterance with respect to the current UBM model. This is done iteratively.
  2. Building VTLN-UBM: A new VTLN-UBM model is built using the VTLN-warped utterances.
  3. Iteration: Steps 1 and 2 are repeated for a few iterations to refine the VTLN-UBM.

For speaker identification, speaker-specific models are then adapted from this VTLN-UBM using an iterative Maximum A Posteriori (MAP) adaptation technique. The paper compares the performance of models adapted from this VTLN-UBM against those adapted from a conventional GMM-UBM.

Datasets

The authors evaluated their proposed method on three standard speech corpora:

  • MIT: Used for adaptation and testing, with a portion used for building the UBM.
  • TIMIT: Used for adaptation and testing, with a portion used for building the UBM.
  • NIST 2004 SRE: Used for evaluation, with the UBM trained on data from NIST 2002 SRE (cell & landline) and Switchboard-1 Release-2 (landline) databases.

Key Insights and Ideas

  • VTLN for UBM Compactness: The central idea is that VTLN, by reducing inter-speaker spectral variability, can lead to a more compact UBM. This compactness is key to improving speaker separability.
  • Improved Speaker Model Separability: A more compact UBM results in speaker-adapted models that are better separated in the acoustic space, meaning they are less likely to overlap with models of other speakers. This is visualized by higher likelihood ratios between the true speaker and competing speakers.
  • MAP Adaptation Dynamics: The VTLN-UBM, being more compact, might require more MAP adaptation iterations to converge effectively compared to a GMM-UBM. The first iteration might not show significant gains, but subsequent iterations yield substantial improvements.
  • Trade-off in Initial Likelihood: While VTLN-UBM adaptation ultimately leads to better results, the initial likelihood of adaptation utterances with respect to the unadapted VTLN-UBM can be lower than with a GMM-UBM. This is analogous to how un-warped utterances might perform worse with VTLN-HMMs in speech recognition.

Key Results

  • Reduced Identification Error Rates (IER): Across all tested databases (NIST 2004 SRE, MIT, TIMIT), the VTLN-UBM system consistently achieved lower Identification Error Rates (IER) compared to the conventional GMM-UBM system.
  • Significant Improvement with Multiple Iterations: For NIST 2004 SRE, using multiple MAP iterations (specifically the second iteration) showed significant performance improvements for the VTLN-UBM system. For example, with 128-component models, IER was reduced after the second iteration.
  • Better N-Best Performance: In N-Best identification tasks (where the system ranks the top N most likely speakers), VTLN-UBM consistently outperformed GMM-UBM across different values of N for NIST 2004 SRE, MIT, and TIMIT.
  • Qualitative Likelihood Analysis: Figures illustrate that adapted speaker models from VTLN-UBM exhibit better separability, with higher likelihood ratios between the correct speaker and the nearest competing speaker.

Limitations and Future Work

  • MAP Iteration Dependency: The paper notes that VTLN-UBM might require more MAP adaptation iterations for optimal performance due to its compactness. This suggests a potential trade-off in adaptation time or complexity.
  • Future Work: The authors' findings suggest that further exploration of VTLN-UBM with different adaptation strategies and larger datasets could yield even better results. The paper implicitly points towards the benefit of exploring more sophisticated adaptation techniques that can leverage the compactness of the VTLN-UBM.

Summary updated

BibTeX
@inproceedings{sarkar2009text,
  author = {A. K. Sarkar and S. Umesh and S. P. Rath},
  title = {{Text-independent speaker identification using vocal tract length normalization for building universal background model}},
  booktitle = {Interspeech},
  year = {2009},
}
Figure